Multi-dimensional characteristic analysis and fault discrimination method for recording signals of transformer substation

By constructing a dual-channel MLP-Attention model and using the Grey Wolf optimization algorithm to optimize the transformer fault identification model, the problem of low accuracy in transformer fault identification in the existing technology is solved, and higher identification accuracy and anti-interference ability are achieved.

CN118937888APending Publication Date: 2024-11-12SKILL TRAINING CENT OF STATE GRID JIANGSU ELECTRIC POWER CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410956827.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing transformer fault identification models are insufficient in capturing the random characteristics and potential fault modes of transformer faults, resulting in low identification accuracy.

Method used

A dual-channel MLP-Attention model was constructed, and the number of hidden layers and hidden layer nodes of the model were optimized using the Grey Wolf Optimization Algorithm. The single-channel MLP model was then optimized by combining the attention mechanism to construct a fault detection model.

Benefits of technology

It improves the accuracy and anti-interference capability of transformer fault identification, and is suitable for substation fault recording data analysis and fault review.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118937888A_ABST
    Figure CN118937888A_ABST
Patent Text Reader

Abstract

The invention discloses a transformer station recording signal multi-dimensional characteristic analysis and fault discrimination method comprising the following steps: obtaining transformer station fault recording signal data, and pre-processing the fault recording signal data; constructing a single-channel MLP model, and optimizing the single-channel MLP model by using an attention mechanism; the method comprises the following steps: constructing a dual-channel MLP-Attention model; a grey wolf optimization algorithm is used to optimize parameters of the dual-channel MLP-Attention model, the training set is used to train the optimized dual-channel MLP-Attention model, and the trained dual-channel MLP-Attention model is used as a fault discrimination model; and detecting the fault discrimination model by using the test set to discriminate the fault type of the transformer substation. According to the method, the two-channel MLP-Attention model is constructed, the number of hidden layers and the number of hidden layer nodes of the model are optimized by using the GWO optimization algorithm, and the method has the advantages of high recognition accuracy and strong anti-interference capability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of transformer substation fault identification, and in particular to a transformer substation recording signal multi-dimensional feature analysis and fault identification method. Background Art

[0002] Transformers are important equipment in power systems, but their working environment is remote and harsh, and they may experience faults including insulation aging and short circuits. When a transformer fails, it will not only affect the operation of the power system, but also cause serious losses to personnel or equipment. Therefore, accurately identifying transformer faults is crucial to ensuring the safety of the power system.

[0003] With the development of artificial intelligence technology, machine learning has become an important means of transformer fault identification. For example, support vector machines, random forests, multi-layer perceptrons, Bayesian methods, etc. have been widely used. Although heuristic algorithms can be applied in the field of power systems, transformer faults have complex and changeable characteristics, such as temperature anomalies and partial discharges, which makes the existing identification models insufficient in capturing these random features and potential fault modes, resulting in low fault identification accuracy. Summary of the invention

[0004] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a method for multi-dimensional feature analysis and fault identification of substation recording signals.

[0005] Technical solution: A method for multi-dimensional feature analysis and fault identification of substation recording signals of the present invention comprises the following steps:

[0006] Obtaining substation fault recording signal data, preprocessing the fault recording signal data, and dividing the preprocessed fault recording signal data into a training set and a test set in proportion;

[0007] Construct a single-channel MLP model, use the attention mechanism to optimize the single-channel MLP model, and record the optimized MLP model as the MLP-Attention model;

[0008] Use the single-channel MLP model and the MLP-Attention model to build a dual-channel MLP-Attention model;

[0009] The parameters of the dual-channel MLP-Attention model are optimized using the Gray Wolf Optimization Algorithm, and the optimized dual-channel MLP-Attention model is trained using the training set. The trained dual-channel MLP-Attention model is used as a fault discrimination model. The input of the fault discrimination model is the substation fault recording signal data, and the output is the substation fault type.

[0010] The test set is used to test the fault discrimination model and to discriminate the fault type of the substation.

[0011] Furthermore, the step of preprocessing the fault recording signal data includes:

[0012] After Fourier transforming the abc three-phase fault recording signal data, the DC components are taken as the first characteristic component, the second characteristic component and the third characteristic component respectively, and the abc three-phase energy values ​​of the fault recording signal data are taken as the fourth characteristic component, the fifth characteristic component and the sixth characteristic component respectively.

[0013] Furthermore, the process of optimizing the single-channel MLP model using the attention mechanism includes:

[0014] The attention mechanism is placed between the input layer and the hidden layer of the single-channel MLP model, and the attention mechanism is used to select different weights for the input features of the single-channel MLP model.

[0015] Furthermore, the parameters of the dual-channel MLP-Attention model include the number of hidden layers and the number of hidden layer nodes of the single-channel MLP model, and the number of hidden layers and the number of hidden layer nodes of the MLP-Attention model.

[0016] Furthermore, the process of optimizing the parameters of the dual-channel MLP-Attention model using the Gray Wolf Optimization Algorithm includes the following steps:

[0017] Step 401, using the parameters of the dual-channel MLP-Attention model as the initial population individuals, initializing the gray wolf population and gray wolf positions, and setting the maximum number of iterations;

[0018] Step 402, calculate the fitness function value of each individual gray wolf; wherein the mean square error MSE is used as the fitness function, and the calculation formula is:

[0019]

[0020] Where Y i is the actual value of the fault category, is the predicted value of the fault category, n is the total number of predictions;

[0021] Step 403, according to the fitness function value, save the top three best gray wolves as the optimal solution α, the second best solution β and the third best solution δ;

[0022] Step 404, updating the position of the ω wolf according to the α wolf, the β wolf and the δ wolf, and updating the parameters A, C and a;

[0023] The position update formula of ω wolf is:

[0024]

[0025] In the formula, z(t+1) represents the adjusted position of ω wolf, t represents the number of iterations, z1, z2, and z3 represent the updated positions of α wolf, β wolf, and δ wolf respectively, and the calculation formulas are:

[0026] z1=z α -A1Dα

[0027] z2=z β -A2D β

[0028] z3=z δ -A3D δ

[0029] In the formula, zα, z β , zδ respectively represent the positions of α wolf, β wolf, and δ wolf before updating, A1, A2, A3 respectively represent the coefficient vectors of α, β, and δ; D α , D α , D α Respectively represent the distance between α wolf, β wolf, δ wolf and prey, and the calculation formula is:

[0030] D α =|C1·z α -z|

[0031] D β =|C2·z β z|

[0032] D δ =|C3·z δ -z|

[0033] In the formula, C1, C2, and C3 are random parameters, and z represents the position distance between ω wolf and α wolf;

[0034] The update formula for parameters A, C and a is:

[0035] A=2a(r1-1)

[0036] C=2r2

[0037]

[0038] Where A is the convergence factor, C is the swing factor, a is a constant that decreases linearly from 2 to 0, constants r1, r2∈(0,1), t max is the maximum number of iterations;

[0039] Step 405, calculating the fitness function values ​​of all gray wolves, and updating the fitness function values ​​and positions of α wolf, β wolf and δ wolf;

[0040] Step 406, determine whether the maximum number of iterations has been reached, if not, go to step 404; if so, end the iteration and obtain the position information of the α wolf, so as to determine the parameters of the dual-channel MLP-Attention model, including the number of hidden layers and hidden layer nodes of the single-channel MLP model, and the number of hidden layers and hidden layer nodes of the MLP-Attention model.

[0041] Furthermore, in the dual-channel MLP-Attention model, the activation function of the single-channel MLP model is the Sigmoid function, and the activation function of the MLP-Attention model is the ReLU function.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0043] The present invention constructs a dual-channel MLP-Attention model and uses the GWO optimization algorithm to optimize the number of hidden layers and hidden layer nodes of the model. Compared with the traditional model, the present model has the advantages of high recognition accuracy and strong anti-interference ability, and is suitable for substation fault recording data analysis, fault reconstructing and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a method for multi-dimensional feature analysis and fault identification of substation recording signals;

[0045] Figure 2 It is a flowchart of a method for multi-dimensional feature analysis and fault identification of substation recording signals;

[0046] Figure 3 It is a schematic diagram of the structure of a single-channel MLP model;

[0047] Figure 4 Optimization diagram of a single-channel MLP model;

[0048] Figure 5 It is a schematic diagram of the structure of the dual-channel MLP-Attention model;

[0049] Figure 6 It is the GWO fitness curve;

[0050] Figure 7 is the classification confusion matrix. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0052] The flowchart of the multi-dimensional feature analysis and fault identification method of substation recording signal described in this embodiment is as follows: Figure 1-2 As shown, the method comprises the following steps:

[0053] Obtaining substation fault recording signal data, preprocessing the fault recording signal data, and dividing the preprocessed fault recording signal data into a training set and a test set in proportion;

[0054] Construct a single-channel MLP model, use the attention mechanism to optimize the single-channel MLP model, and record the optimized MLP model as the MLP-Attention model;

[0055] Use the single-channel MLP model and the MLP-Attention model to build a dual-channel MLP-Attention model;

[0056] The parameters of the dual-channel MLP-Attention model are optimized using the Gray Wolf Optimization Algorithm, and the optimized dual-channel MLP-Attention model is trained using the training set. The trained dual-channel MLP-Attention model is used as a fault discrimination model. The input of the fault discrimination model is the substation fault recording signal data, and the output is the substation fault type.

[0057] The test set is used to test the fault discrimination model and to discriminate the fault type of the substation.

[0058] Specifically, the steps of preprocessing the fault recording signal data include:

[0059] After Fourier transform of the abc three-phase fault recording signal data, the DC component is taken as the first characteristic component, the second characteristic component and the third characteristic component respectively, and the abc three-phase energy value of the fault recording signal data is taken as the fourth characteristic component, the fifth characteristic component and the sixth characteristic component respectively, where the energy value is obtained by adding the absolute value of the fault recording sampling value and dividing it by the sampling number.

[0060] Specifically, the process of optimizing the single-channel MLP model using the attention mechanism includes:

[0061] The attention mechanism is placed between the input layer and the hidden layer of the single-channel MLP model, and the attention mechanism is used to select different weights for the input features.

[0062] like Figure 3 As shown, the existing single-channel MLP model (multi-layer perceptron) includes an input layer, multiple hidden layers and an output layer, x1, x2, ..., x n As input features, y1, y2, ..., y n As output features. Figure 4As shown, in this embodiment, the attention mechanism is placed between the input layer and the hidden layer of the MLP model. During the training process, the attention mechanism automatically increases the weight of important input features and ignores unimportant input features. The input features after weight adjustment are then input into the hidden layer, where e1, e2, ..., e n The input feature is obtained by performing a tanh operation, and then the features α1, α2, ..., α are obtained by performing an activation function Softmax. n , the final output features are H1, H2, ..., H n , and input it into the hidden layer. In this embodiment, the single-channel MLP model is optimized by using the attention mechanism, and the optimized MLP model is recorded as the MLP-Attention model.

[0063] For a single-channel MLP model, after its hidden layer receives the input feature vector, it passes it to the first hidden layer. The hidden layer includes multiple hidden units. For the first hidden layer, the output value of hidden unit j is calculated using the following formula:

[0064]

[0065] For the Lth hidden layer, calculate the output value of hidden unit j, and the calculation formula is:

[0066]

[0067] For the output layer, the output value of output unit k is calculated using the following formula:

[0068]

[0069] In the formula, a j is the output value of the first hidden unit, W ij is the weight from the input layer to the hidden layer, x i The input layer is the input, n is the number of input features, b j is the bias of the hidden layer; a p is the output value of the Lth hidden unit, p is the number of neurons in the L-1th hidden layer, W jp is the weight from the L-1th hidden layer to the Lth hidden layer, a j is the input of the Lth hidden layer, b p is the bias of the Lth hidden layer; y k is the output value of the output unit, W pk is the weight from the hidden layer to the output layer, a p is the output of the hidden layer, b k is the bias of the output layer.

[0070] like Figure 5As shown in the figure, the dual-channel MLP-Attention model includes a single-channel MLP model and an MLP-Attention model. In the dual-channel MLP-Attention model, the activation function of the single-channel MLP model is the Sigmoid function, and the activation function of the MLP-Attention model is the ReLU function. The feature N is obtained through the single-channel MLP model, and the feature M is obtained through the MLP-Attention model. The two are weighted to obtain the output feature Y of the dual-channel MLP-Attention model. The calculation formula is:

[0071] Y=β1·M+β2·N

[0072] In the formula, β1 and β2 are weighted coefficients; the calculation formulas of M and N are:

[0073] M = ReLU(X att w h1 +b h1 )

[0074] X att =X☉α

[0075] N=Sigmoid(Xw h2 +b h2 )

[0076] In the formula, w h1 、w h2 is the weight matrix, b h1 , b h2 is the bias, X att is the output of the attention mechanism, and X is the input of the attention mechanism module.

[0077] The parameters of the dual-channel MLP-Attention model include the number of hidden layers and hidden layer nodes of the single-channel MLP model, and the number of hidden layers and hidden layer nodes of the MLP-Attention model. The Grey Wolf Optimization (GWO) algorithm is used to optimize the number of hidden layers and hidden layer nodes of the single-channel MLP model, and the number of hidden layers and hidden layer nodes of the MLP-Attention model.

[0078] The process of optimizing the parameters of the dual-channel MLP-Attention model using the Gray Wolf Optimization Algorithm includes the following steps:

[0079] Step 401, using the parameters of the dual-channel MLP-Attention model as the initial population individuals, initializing the gray wolf population and gray wolf positions, and setting the maximum number of iterations;

[0080] Step 402, calculate the fitness function value of each individual gray wolf; wherein the mean square error MSE is used as the fitness function, and the calculation formula is:

[0081]

[0082] Where Y i is the actual value of the fault category, is the predicted value of the fault category, n is the total number of predictions;

[0083] Step 403, according to the fitness function value, save the top three best gray wolves as the optimal solution α, the second best solution β and the third best solution δ;

[0084] Step 404, updating the position of wolf ω according to α, β and δ, and updating parameters A, C and a;

[0085] The position update formula of ω wolf is:

[0086]

[0087] In the formula, z(t+1) represents the adjusted position of ω wolf, t represents the number of iterations, z1, z2, z3 represent the updated positions of α, β, δ wolf respectively, and the calculation formulas are:

[0088] z1=z α -A1D α

[0089] z2=z β -A2D β

[0090] z3=z δ -A3D δ

[0091] In the formula, z α 、z β 、z δ They represent the positions of α, β, and δ wolves before the update, respectively. A1, A2, and A3 represent the coefficient vectors of α, β, and δ, respectively. D α , D α , D α Respectively represent the distance between the wolf and the prey, α, β, δ, and the calculation formula is:

[0092] Dα=|C1·z α -z|

[0093] D β =|C2·z β -z|

[0094] D δ =|C3·z δ -z|

[0095] In the formula, C1, C2, and C3 are random parameters, and z represents the position distance between ω wolf and α wolf;

[0096] The update formula for parameters A, C and a is:

[0097] A=2a(r1-1)

[0098] C=2r2

[0099]

[0100] Where A is the convergence factor, C is the swing factor, a is a constant that decreases linearly from 2 to 0, and the constants r1, r2∈(0,1); t max is the maximum number of iterations;

[0101] Step 405, calculating the fitness function values ​​of all gray wolves, and updating the fitness function values ​​and positions of α wolf, β wolf and δ wolf;

[0102] Step 406, determine whether the maximum number of iterations has been reached, if not, go to step 404; if so, end the iteration and obtain the position information of the α wolf, so as to determine the parameters of the dual-channel MLP-Attention model, including the number of hidden layers and hidden layer nodes of the single-channel MLP model, and the number of hidden layers and hidden layer nodes of the MLP-Attention model.

[0103] In one example, the Gray Wolf optimizes the parameters of the dual-channel MLP-Attention model using the mean square error MSE as the fitness function. The fitness curve is as follows: Figure 6 As shown in the figure, after the 12th iteration, the fitness value reaches the minimum, and the parameters of the dual-channel MLP-Attention model are determined according to the number of hidden layers and the number of hidden layer nodes optimized at this time.

[0104] In the example, the acquired substation fault recording signal is preprocessed, and the DC component and energy of the Fourier transform of the obtained three-phase voltage are used as features 1 to 6. Some of the features are shown in Table 1.

[0105] Table 1 Some fault characteristics

[0106]

[0107] The preprocessed data is divided into a training set and a test set according to an 8:2 ratio. The training set is used to train the dual-channel MLP-Attention model, and the test set is used to test the performance of the fault discrimination model. The data set division is shown in Table 2.

[0108] Table 2 Dataset division

[0109]

[0110] After the dual-channel MLP-Attention model is trained, the fault discrimination model is tested using the test set data. The test results are as follows: Figure 7 shown.

[0111] To verify the performance of this method, 30 experiments were conducted to compare the four models of SVM (support vector machine), single-channel MLP, single-channel MLP-Attention model and dual-channel MLP-Attention proposed in this embodiment. The accuracy Acc, precision P, recall R, F 1-score Four indicators are compared, and the calculation formulas are:

[0112]

[0113]

[0114] In the formula, TP refers to the number of samples that are actually positive and predicted to be positive; FN refers to the number of samples that are actually positive and predicted to be negative; FP refers to the number of samples that are actually negative and predicted to be positive; TN refers to the number of samples that are actually negative and predicted to be negative. The comparison results are shown in Table 3.

[0115] Table 3 Comparison results

[0116]

[0117] The results show that the dual-channel MLP-Attention model proposed in this embodiment outperforms other models in all four evaluation indicators, has good performance in transformer fault identification, and has a higher accuracy in substation fault identification.

Claims

1. A method for multi-dimensional feature analysis and fault identification of substation recording signals, characterized in that: The steps include: Obtaining substation fault recording signal data, preprocessing the fault recording signal data, and dividing the preprocessed fault recording signal data into a training set and a test set in proportion; Construct a single-channel MLP model, use the attention mechanism to optimize the single-channel MLP model, and record the optimized MLP model as the MLP-Attention model; Use the single-channel MLP model and the MLP-Attention model to build a dual-channel MLP-Attention model; The parameters of the dual-channel MLP-Attention model are optimized using the Gray Wolf Optimization Algorithm, the optimized dual-channel MLP-Attention model is trained using the training set, and the trained dual-channel MLP-Attention model is used as a fault discrimination model. The input item of the fault identification model is the substation fault recording signal data, and the output item is the substation fault type; The test set is used to test the fault discrimination model and to discriminate the fault type of the substation.

2. A method for multi-dimensional feature analysis and fault identification of substation recording signals according to claim 1, characterized in that: The steps of preprocessing the fault recording signal data include: After Fourier transforming the abc three-phase fault recording signal data, the DC components are taken as the first characteristic component, the second characteristic component and the third characteristic component respectively, and the abc three-phase energy values ​​of the fault recording signal data are taken as the fourth characteristic component, the fifth characteristic component and the sixth characteristic component respectively.

3. A method for multi-dimensional feature analysis and fault identification of substation recording signals according to claim 1, characterized in that: The process of optimizing a single-channel MLP model using the attention mechanism includes: The attention mechanism is placed between the input layer and the hidden layer of the single-channel MLP model, and the attention mechanism is used to select different weights for the input features of the single-channel MLP model.

4. A method for multi-dimensional feature analysis and fault identification of substation recording signals according to claim 3, characterized in that: The parameters of the dual-channel MLP-Attention model include the number of hidden layers and hidden layer nodes of the single-channel MLP model, and the number of hidden layers and hidden layer nodes of the MLP-Attention model.

5. A method for multi-dimensional feature analysis and fault identification of substation recording signals according to claim 4, characterized in that: The process of optimizing the parameters of the dual-channel MLP-Attention model using the Gray Wolf Optimization Algorithm includes the following steps: Step 401, using the parameters of the dual-channel MLP-Attention model as the initial population individuals, initializing the gray wolf population and gray wolf positions, and setting the maximum number of iterations; Step 402, calculate the fitness function value of each individual gray wolf; wherein the mean square error MSE is used as the fitness function, and the calculation formula is: Where Y i is the actual value of the fault category, is the predicted value of the fault category, n is the total number of predictions; Step 403, according to the fitness function value, save the top three best gray wolves as the optimal solution α, the second best solution β and the third best solution δ; Step 404, updating the position of the ω wolf according to the α wolf, the β wolf and the δ wolf, and updating the parameters A, C and a; The position update formula of ω wolf is: In the formula, z(t+1) represents the adjusted position of ω wolf, t represents the number of iterations, z1, z2, and z3 represent the updated positions of α wolf, β wolf, and δ wolf respectively, and the calculation formulas are: z1=z α -A1D α z2=z β -A2D β <h2 style=";text-align:left;direction:ltr">z3=z<h2 style=";text-align:left;direction:ltr"> δ <h2 style=";text-align:left;direction:ltr"> -A3D<h2 style=";text-align:left;direction:ltr"> δ In the formula, z α 、z β 、z δ They represent the positions of α wolf, β wolf, and δ wolf before updating, respectively; A1, A2, and A3 represent the coefficient vectors of α, β, and δ, respectively; D α , D α , D α Respectively represent the distance between α wolf, β wolf, δ wolf and prey, and the calculation formula is: D α =|C1·z α -z| D β =|C2 with β -z| D δ =|C3·z δ -z| In the formula, C1, C2, and C3 are random parameters, and z represents the position distance between ω wolf and α wolf; The update formula for parameters A, C and a is: A=2a(r1-1) C=2r2 Where A is the convergence factor, C is the swing factor, a is a constant that decreases linearly from 2 to 0, constants r1, r2∈(0,1), t max is the maximum number of iterations; Step 405, calculating the fitness function values ​​of all gray wolves, and updating the fitness function values ​​and positions of α wolf, β wolf and δ wolf; Step 406, determine whether the maximum number of iterations has been reached, if not, go to step 404; if so, end the iteration and obtain the position information of the α wolf, so as to determine the parameters of the dual-channel MLP-Attention model, including the number of hidden layers and hidden layer nodes of the single-channel MLP model, and the number of hidden layers and hidden layer nodes of the MLP-Attention model.

6. A method for multi-dimensional feature analysis and fault identification of substation recording signals according to claim 1, characterized in that: In the dual-channel MLP-Attention model, the activation function of the single-channel MLP model is the Sigmoid function, and the activation function of the MLP-Attention model is the ReLU function.